Google reports Antigravity /teamwork results on open math and CS problems

On November 13, 2026, Vahab Mirrokni, Google Fellow and VP at Google Research, published a post on results from Antigravity's /teamwork multi-agent systems. Over several weeks, teams at Google and academic partners used Gemini 4 Argon, Gemini 4 Carbon, Gemini 4 Hydrogen and /teamwork workflows on problems that had stayed open for years, in mathematics, theoretical computer science, chemistry, systems software and ML engineering. Google stresses that all of the work was done with human experts, and says it is making the tools available in Antigravity's /teamwork, alongside /boost and other agents.

The design rests on two stated limits of current models. First, generation is unpredictable, so /teamwork pairs parallel exploration with dedicated falsifiers (the Lean 4 proof checker, unmodified software test suites, execution simulators, web search, hardware counters) to verify automatically which candidates work. Second, models lose track of details over multi-day tasks as context grows, so /teamwork partitions context: each specialized agent gets only the files and tools needed for a step, coordinators hand off to fresh successors, and critics and auditors check results from clean contexts without the producer's reasoning. Workflows are written as declarative, self-improving natural-language patterns, and each user prompt is routed to the best-suited one. Patterns range from distributed coding and iterative problem-solving (packaged standalone as /boost, a sequential solve-and-verify loop) to long proof and document review. Many are available in Antigravity now, with more updates promised in the coming weeks.

The headline result is the 2013 Courtade-Kumar conjecture in information theory, proved in full generality by researchers from Google, CUHK, USC and CMU using /teamwork's Long Proof pattern (implementing Stellar Colosseum in Antigravity). The conjecture says that when a random bit string passes through a noisy channel, no compression of the input to a single bit preserves more information than a 'dictator function' that copies one input bit. Gemini first found counterexamples to a generalized conjecture from the CUHK team's earlier paper. The CUHK team then proposed a new candidate function that reduced the problem to a 14-variable optimization problem, plus a key analytical lower bound. Gemini 4 Argon and earlier Gemini models partitioned the parameter space and developed most of the new analytical bounds and interval-arithmetic certificates, and extended the proof from balanced to unbalanced functions without human input. The whole proof was formalized in Lean 4, producing 50 million lines of code.

Other theory results came from posing open problems directly to Gemini 4 and to harnesses such as Stellar Colosseum and /teamwork; the authors verified, simplified and edited the outputs for publication. A harness on Gemini 4 Argon proved and Lean-verified near-optimal sample bounds for adaptively chosen statistical queries, closing a gap open since 2014, which yields the first nearly space-optimal streaming algorithm for data with insertions and deletions that stays accurate against adaptive adversaries. In quantum computing, a new construction found the first n-bit Boolean functions needing strictly more than n/2 quantum queries at constant error, settling a 2014 question. For fifty years until STOC 2024, combinatorial algorithms for Boolean matrix multiplication and triangle detection saved only logarithmic factors over cubic time; a new deterministic algorithm resolves STOC 2024's main open question. Stacked sparse sketches resolve a 2025 question in randomized linear algebra with only √log r nonzeros per row, quadratically fewer than the log r needed to preserve all lengths. A new lower bound shows that approximating all pairwise distances among n points within any constant factor can require polynomially more than n edges, even with Steiner points and arbitrarily many hops. Stellar Colosseum also produced the first single-pass streaming algorithm for cut sparsification of weighted graphs with near-optimal memory, update time and reconstruction time, resolving a FOCS 2022 question.

Further items: a deterministic algorithm for online inverse linear optimization with optimal regret that cuts runtime from exponential in the dimension to polynomial; multi-agent campaigns that simplified the theoretical Turing-machine bound for n-bit integer multiplication below n log n; a separate new O(n log n) algorithm on a single two-dimensional grid whose C implementation matches GMP bit-for-bit up to a billion bits; significant progress (not a full resolution) on Tarjan's 1985 Deque Conjecture, with a new upper bound verified in Lean 4 that beats every finitely iterated inverse-Ackermann function; and a proof that approximate leave-one-out cross-validation stays accurate for non-smooth sparse penalties such as elastic net.

The post explains why Lean matters: checking a computer-assisted manuscript by hand burdens reviewers, while a Lean 4 formalization is checked line by line by a kernel. Once human mathematicians confirm the formal statement matches the original problem, the proof compiles only if every logical step holds. Working from informal proofs, Google's multi-agent framework LEAP, developed with colleagues in Google Cloud AI and paired with Argon, produced several formalizations. /teamwork-LEAP fully autonomously produced an end-to-end, machine-checked formalization of the Weil conjectures (culminating in Pierre Deligne's 1974 proof), about 1.8 million lines of Lean across 28,800 subagent invocations. It synthesized over 919,000 lines of Lean for the first formalization of a recent proof of Hilbert's tenth problem over general rings of integers. It independently reproduced an end-to-end formalization of a recent proof of finite-time blowup for the 3D Navier-Stokes equations with smooth forcing, part of one of the seven Millennium Prize Problems and open for about 90 years. It also generated the first formalization of the Birkhoff conjecture (open since 1927) and of the 3D Bochner-Riesz conjecture (open for more than 50 years), finding and fixing a gap in a recently released informal proof along the way. Other machine-checked results include the C1 regularity of infinity-harmonic functions in three and higher dimensions (open for about 20 years), Killip and Vişan's theorem on KdV well-posedness in H^-1, and verification of a UCLA group's informal proofs of the 40-year-old Jamison Caterpillar Conjecture, the Ağırseven-Ollis conjecture and an improved upper bound on Erdős Problem #1063.

Google also offered Deep Review to authors ahead of STOC 2027: authors of 578 submissions, over a third of all submissions, opted in for feedback on drafts. In an ongoing survey (N = 70), 90% rated the feedback mostly or almost all factual and 90% would use it again. Over 30% said it caught substantive proof bugs or mathematically significant typos, and 14% said it found technical gaps that took more than an hour to fix. Users get two free Deep Review calls per week, and it is also available as /teamwork's document review pattern in Antigravity.

The post then turns to laboratory science, saying that pairing generator agents with independent verifier agents enables discovery workflows. Its first example is determining a molecule's structure from nuclear magnetic resonance data, where, in HIV protease inhibitors, one misplaced bond separates an inactive look-alike from a drug, and the work organizes 30 specialized agents. The post also mentions kernel optimization and several production infrastructure gains at Google; the details of the chemistry, systems software and ML engineering results are not covered in this retelling.

Key facts

  • Google Research says /teamwork, on Gemini 4 Argon, Carbon and Hydrogen and with human experts, helped prove the 2013 Courtade-Kumar conjecture in full generality; the Lean 4 formalization runs to 50 million lines of code.
  • /teamwork-LEAP is said to have fully autonomously produced a machine-checked Lean formalization of the Weil conjectures: about 1.8 million lines across 28,800 subagent invocations.
  • The system pairs parallel exploration with falsifiers such as the Lean 4 checker and test suites, and gives each agent only the files and tools it needs for a step.
  • Deep Review drew opt-ins from authors of 578 STOC 2027 submissions; in an ongoing survey (N = 70), 90% rated the feedback mostly or almost all factual.
  • Many features are in Antigravity now, with more for external users promised in the coming weeks; Deep Review gives two free calls per week.

Why it matters

The post is Google's case that multi-agent harnesses with automatic checking can contribute to research-level work. The strongest claims rest on verification that does not depend on trusting the model: Lean 4 proofs that compile only if every step holds. Reported outputs include a full-generality proof of the Courtade-Kumar conjecture, a Lean formalization of the Weil conjectures of about 1.8 million lines, and closed gaps in adaptive data analysis (open since 2014) and quantum query complexity (a 2014 question). It also resolves STOC 2024's main open question on combinatorial Boolean matrix multiplication and triangle detection. The tools behind these results are being offered to outside researchers and developers.

Who it affects

Mathematicians and theoretical computer scientists are the main audience, since most results are proofs and formalizations. Authors submitting to conferences are affected through Deep Review, which was offered ahead of STOC 2027. Researchers in chemistry and engineers working on kernel optimization and systems software are also named as targets, with Google citing production infrastructure gains internally. Peer reviewers could be affected too: the post argues that Lean formalization eases the burden of checking computer-assisted manuscripts by hand.

How to use it

Google says the agentic tools are available in Antigravity's /teamwork, alongside /boost, which packages a sequential solve-and-verify loop. /teamwork routes each prompt to the pattern best suited to it; the patterns include distributed coding, iterative problem-solving, long proof and document review, and the post says all of these are available in Antigravity today. Teams can also design custom specifications from past run transcripts. Deep Review is available as /teamwork's document review pattern, with two free calls per week. Google says it will provide more updates in the coming weeks covering all features for external users. No pricing or usage limits for /teamwork or /boost are given.

How solid is it

This is a first-party post by Google Research, and the results are self-reported. The strongest evidence offered is machine checking in Lean 4, which the post says is trustworthy once human mathematicians confirm the formal statement matches the original problem. The post does not say any of the results are peer reviewed or accepted for publication; it says the authors verified, simplified and edited the outputs for publication. Several items are reproductions or formalizations of existing proofs, such as the Navier-Stokes blowup result, which reproduces a recent proof rather than presenting a new one. The Deep Review survey is ongoing and small: N = 70 against 578 opted-in submissions. The post does not give benchmark comparisons of the three Gemini 4 models or describe how they differ.

Risks and caveats

Credit is shared: Google says all work was done with human experts, and for Courtade-Kumar the CUHK team proposed the new candidate function and a key analytical lower bound. Gemini is credited with developing most of the new analytical bounds and interval-arithmetic certificates, so it is not a story of an unaided model. A Lean proof is only as meaningful as the formal statement, which humans must confirm matches the original problem. The Tarjan Deque Conjecture result is significant progress with a new upper bound, not a full resolution. Many features are not yet available externally, and no timeline beyond the coming weeks is given. Survey respondents who found the tool useful are a self-selected group of authors who opted in.

“Tools of this kind can help produce cleaner and more precise research, faster”

— Surveyed author (unnamed), Deep Review survey, quoted in Google Research's post